Using sliced inverse mean difference for dimension reduction in multivariate time series

Following recent developments of dimension reduction algorithms for a multivariate time series, we propose in this work the adaptation of sliced inverse mean difference algorithm, an algorithm which was previously proposed in a standard multiple regression setting, to develop an algorithm appropriate to perform dimension reduction for a multivariate time series. The resulting algorithm called time series sliced inverse mean difference (TSIMD) is shown to be able to identify important directions and important lags using less significant pairs than previously proposed algorithms for dimension reduction in multivariate time series. We demonstrate the competitive performance of our algorithms through a number of experiments.